Anthropic’s Claude 3.5 has been getting a lot of attention in enterprise circles lately. The marketing claims are impressive – better reasoning, longer context windows, improved safety. But what’s the actual story on the ground?

I’ve spent the past month talking to CIOs and technology leaders across Australian enterprises about their Claude experiences. The picture is more nuanced than either the hype or the scepticism suggests.

The Adoption Numbers

Let’s start with what we can observe. Among the forty-odd enterprise technology leaders I surveyed informally:

• About 35% have active Claude pilots running
• Another 40% are evaluating or planning evaluations
• The remaining 25% are sticking with GPT-4 or waiting

That’s higher adoption than I would have predicted six months ago. Something has shifted.

What’s Driving Interest

Three factors keep coming up in conversations:

The context window. Claude 3.5’s ability to handle longer documents is genuinely useful for enterprise use cases. Legal document review, code analysis, research synthesis – these all benefit from being able to process more content without chunking.

Safety and predictability. Anthropic’s constitutional AI approach resonates with compliance-conscious enterprises. Several CIOs mentioned feeling more comfortable putting Claude in front of internal users because it’s less likely to produce unexpected outputs.

Pricing competition. Having a credible alternative to OpenAI is creating pricing pressure. Even enterprises that don’t switch are using Claude evaluations as negotiating leverage.

Where It’s Working

The most successful Claude deployments I’ve seen share some characteristics:

Document-heavy workflows. Contract analysis, policy review, research summarisation. The extended context window shines here.

Internal tools first. Companies are deploying Claude for internal use before customer-facing applications. Lower risk, faster iteration.

Augmentation rather than automation. The successful projects treat Claude as a tool for professionals, not a replacement for them. A lawyer using Claude to review contracts faster, not Claude replacing lawyers.

Where It’s Struggling

Not everything is rosy:

Integration complexity. Swapping one LLM for another isn’t trivial. Prompt engineering that works for GPT-4 doesn’t automatically transfer. Teams are spending significant effort on adaptation.

Ecosystem gaps. OpenAI’s ecosystem is more mature. More tutorials, more third-party tools, more community knowledge. Claude adopters are often figuring things out from scratch.

Consistency concerns. Several teams reported more variability in Claude outputs than they expected. The model performs brilliantly on some tasks and puzzlingly poorly on similar ones. This unpredictability is challenging for production systems.

The Australian Context

A few Australia-specific observations:

Data sovereignty matters. Anthropic’s infrastructure story is less clear than Azure’s for Australian compliance requirements. Some enterprises can’t proceed until this is resolved.

Local expertise is thin. Finding consultants and contractors with deep Claude experience is harder than finding GPT-4 expertise. The talent pool hasn’t caught up.

Wait-and-see is common. Australian enterprises tend to be more conservative than US counterparts. Many are watching early adopters before committing.

What I’d Recommend

If you’re evaluating Claude 3.5 for enterprise use:

1. Start with a specific use case. Don’t try to replace your entire AI stack at once. Pick one workflow where Claude’s strengths align with your needs.

2. Budget for adaptation. Expect 2-3 months of engineering work to adapt existing prompts and integrations. It’s not plug-and-play.

3. Run parallel evaluations. Test the same tasks on Claude and GPT-4. Understand where each excels before committing.

4. Check the compliance details. Understand exactly where your data goes and what Anthropic’s commitments are. Don’t assume it’s the same as Azure OpenAI.

5. Plan for model churn. Both Anthropic and OpenAI are releasing updates frequently. Your architecture needs to handle model changes without massive rework.

The Bigger Picture

Claude 3.5 is a legitimate enterprise option. It’s not clearly better than GPT-4 across the board, but it’s clearly better for certain use cases. The days of OpenAI having the only credible enterprise model are over.

That’s good for everyone. Competition improves offerings and keeps pricing reasonable. The practical winner is enterprises who can evaluate both options and choose based on their specific needs rather than vendor lock-in.

The hype cycle will keep cycling. Meanwhile, the real work is figuring out where these tools actually deliver value in your specific context. That’s less exciting than breathless AI announcements, but it’s where the ROI actually lives.